On Tue, Mar 20, 2018 at 12:45:50PM +0100, Vlastimil Babka wrote: > On 03/20/2018 09:54 AM, Aaron Lu wrote: > > Running will-it-scale/page_fault1 process mode workload on a 2 sockets > > Intel Skylake server showed severe lock contention of zone->lock, as > > high as about 80%(42% on allocation path and 35% on free path) CPU > > cycles are burnt spinning. With perf, the most time consuming part inside > > that lock on free path is cache missing on page structures, mostly on > > the to-be-freed page's buddy due to merging. > > But why, with all the prefetching in place? The prefetch is just for its order 0 buddy, if merge happens, then its order 1 buddy will also be checked and on and on, so the cache misses are much more in merge mode. > > > One way to avoid this overhead is not do any merging at all for order-0 > > pages. With this approach, the lock contention for zone->lock on free > > path dropped to 1.1% but allocation side still has as high as 42% lock > > contention. In the meantime, the dropped lock contention on free side > > doesn't translate to performance increase, instead, it's consumed by > > increased lock contention of the per node lru_lock(rose from 5% to 37%) > > and the final performance slightly dropped about 1%. > > > > Though performance dropped a little, it almost eliminated zone lock > > contention on free path and it is the foundation for the next patch > > that eliminates zone lock contention for allocation path. > > Not thrilled about such disruptive change in the name of a > microbenchmark :/ Shouldn't normally the pcplists hide the overhead? Sadly, with the default pcp count, it didn't avoid the lock contention. We can of course increase pcp->count to a large enough value to avoid entering buddy and thus avoid zone->lock contention, but that would require admin to manually change the value on a per-machine per-workload basis I believe. > If not, wouldn't it make more sense to turn zone->lock into a range lock? Not familiar with range lock, will need to take a look at it, thanks for the pointer. > > > A new document file called "struct_page_filed" is added to explain > > the newly reused field in "struct page". > > Sounds rather ad-hoc for a single field, I'd rather document it via > comments. Dave would like to have a document to explain all those "struct page" fields that are repurposed under different scenarios and this is the very start of the document :-) I probably should have explained the intent of the document more. Thanks for taking a look at this. > > Suggested-by: Dave Hansen <dave.hansen@xxxxxxxxx> > > Signed-off-by: Aaron Lu <aaron.lu@xxxxxxxxx> > > --- > > Documentation/vm/struct_page_field | 5 +++ > > include/linux/mm_types.h | 1 + > > mm/compaction.c | 13 +++++- > > mm/internal.h | 27 ++++++++++++ > > mm/page_alloc.c | 89 +++++++++++++++++++++++++++++++++----- > > 5 files changed, 122 insertions(+), 13 deletions(-) > > create mode 100644 Documentation/vm/struct_page_field > >
